Bayesian Probabilistic Matrix Factorization — A dive towards recommendation

Abstract

Matrix factorization is a well known technique which discovers latent features among users and items. This method brings the advantage of reducing data sparcity and cold start problem. The different Matrix factorization methods such as SVD, PMF, NMF etc. exists. But all these suffer from certain drawback especially when dataset is extremely sparse. The Bayesian Probabilistic Matrix Factorization (BPMF) method proves to be more efficient and provides prediction that leads to better accuracy. The significance of BPMF is to avoid parameter tuning and provides predictive distribution. To enhance user satisfaction and loyalty particularly when the huge volume of data is available, there is need of recommender system. Hence, the idea of BPMF is extended towards recommendation where top N queries are recommended to users using BPMF method liaison with Cholesky decomposition, Gibbs sampling technique, K nearest neighbor method. The experimental work describes that the BPMF method when used in query recommendation provides better results.